acceptodds
Under review as a conference paper at ICLR 2027

GraphIR: A Canonical Graph Intermediate Representation for Power-Grid Control

Abstract

Graph neural networks are increasingly used for power-grid control, yet Grid2Op observations are typically converted into graphs through model-specific preprocessing. This couples simulator indexing, graph semantics, architecture, and action design, making representation choices difficult to reproduce or compare. We introduce *GraphIR*, a canonical intermediate representation that parses a Grid2Op state once, preserves equipment identity and terminal-level topology, and deterministically projects the state into task-specific graph views. GraphIR exposes representation as a controlled experimental variable while retaining traceability to simulator objects. Our evaluation isolates representation, system, and sampler effects through view ablations, flat and graph baselines, cross-environment validation, runtime measurements, and entity-level diagnostics.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Neural Networks, Reinforcement Learning, Power Systems

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_9LqYOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper aims to present the Grid2Op as graphs to enable power-grid control.

Critical strengths

None, as I hardly can understand this paper, to the very very best of my effort.

Critical weaknesses

The presentation of this paper is far from good enough. So many notations and terms are not introduced/defined before using. Many statements are vague; tables and algorithms are not formally presented; math formulations are not carefully checked. I will list a few points but not all of them will be enumerated because basically this paper requires significant revision.

  1. Fonts are different from other submission
  2. The key term, Grid2Op, is not introduced. I m not an expert on this topic but considering this is a machine-learning conference, such domain-specific terms shall be introduced to broad readers.
  3. Not clear what is “environment specification 𝑆”
  4. Notation C is misused in Eq. 1
  5. Not clear what is “persistent entities 𝑉”
  6. Not clear what is “𝑀 to Grid2Op indices or derived identifiers”
  7. Not clear what does “This is a logical state; its arrays may be materialized lazily.” mean
  8. Table 1’s readability is very bad.
  9. Not clear what does “A projection may subsequently collapse this path, but the canonical state preserves it.” mean
  10. Algorithm 1’s format is not standard and not formal
  11. Eq. 3’s presentation is also terrible: a lot of notations are misplaced at the subscripts of the readout operator.
Questions for authors

Please check the weaknesses.

Reviewer_q4XTOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose GraphIR, an intermediate representation (IR) that parses a Grid2Op observation once (persistent equipment identities, terminal-to-busbar assignment, mappings to Grid2Op indices) and projects it into graph views for RL agents. The paper then compares a two-view GNN (operating + physical views), specifically designed for this IR, against an operating GNN and a flat MLP on Case 14 and Case 118, with an unsafe-state exit test.

Critical strengths
  • The paper is very well written; the IR design is clearly explained, illustrated and well formulated.
  • The framework is rich of implementations: split between static and dynamic data, persistent equipment identities, terminal-to-busbar assignment, mappings to original Grid2Op indices and model agnostic.
  • The authors share negative results explicitly and very honestly: GraphIR doesn't allow better performance than a simple MLP or an operating GNN.
Critical weaknesses
  • Although authors are being humble about it, this is also the main issue. An operating GNN provides better representations than GraphIR (model), for faster. On Case 14 and Case 118, GraphIR is as performant as an MLP. On Figure 5, the error margin is too wide: all intervals contain 0% of improvement, so no conclusion can actually be made at this point.
  • All experiments include the GraphIR representation (Grid2Op observation -> GraphIR TensorDict -> model). The IR is the main contribution of the paper and has no proof showing its efficiency against a direct Grid2Op observation, or other IRs. GraphIR was introduced to tackle/improve reproducibility, comparability and reusability (as presented in the abstract and intro), but no experiment measures it. The idea might be clever, but the claims are not verified at all.
  • On Case 118, it is unclear whether results come from the policy or the action filter. Δρ\Delta\rho is close to 0 (even negative for MLP and GraphIR) but 82-87% of states still exit the danger zone. Unlike Case 14, There is no do-nothing baseline, nor a random pick among the 128 filtered candidates.
  • It is sometimes confusing to see "GraphIR" the representation layer, and "GraphIR" the GNN model. My understanding is that the representation layer is the main contribution, and the model is a proposition that would take fully advantage of the representation. I would suggest the authors to have a solid representation layer that has been tested with existing models first: testing existing graph agents through GraphIR to check that it (at least) matches its results.
Questions for authors
  1. Exit@HH: what's the value of HH? What are τ\tau_act, τ\tau_crit and KK?
  2. In Table 3, given the value of ρ\rho after action and Δρ\Delta\rho, we can see that the pre-action ρ\rho values differ accross policies. Why? Are the 1000 unsafe starts the same for every checkpoint?
  3. Why are Case 14's MLP and GraphIR results identical with zero variance? How many distinct unsafe states are there in the Case 14 experiment?
Reviewer_NguaOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces GraphIR, a canonical intermediate graph representation for power-grid control in Grid2Op. The proposed framework separates simulator-specific observation parsing from downstream graph construction by maintaining persistent equipment identities, terminal-level topology information, and deterministic graph projections. Based on this representation, the authors construct homogeneous and heterogeneous graph views and integrate them into a GNN-based PPO controller. Experiments are conducted on Case 14 and Case 118 environments, comparing GraphIR against a flat MLP and a simpler operating-grid GNN.

Critical strengths

None

Critical weaknesses
  • Lack of novelty and methodological contributions. The proposed GraphIR is essentially a software engineering abstraction that reorganizes existing graph preprocessing operations. Persistent entity identities, typed schemas, and deterministic projections are standard implementation practices rather than meaningful methodological innovations.
  • Unconvincing empirical results. GraphIR barely outperforms the MLP baseline on Case 118 (0.826 vs. 0.822 Exit@H) and underperforms the simpler operating-grid GNN (0.867). The experiments fail to justify the additional complexity.
  • Limited research significance. The work is highly specific to Grid2Op and provides little insight into graph representation learning or reinforcement learning beyond this particular application. Even as an infrastructure contribution, the claimed benefits of reusability and generality are insufficiently demonstrated.
Questions for authors

See above.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.